PHASE 03 · SOP

Master the SOP — The Director's Core Skill

A clear Standard Operating Procedure can make recurring work easier to evaluate and supervise. It is one part of a dependable workflow, alongside permissions, validation, testing, human review, monitoring, and incident response.

⏱ 90–120 min · write 1 real SOP Writing Most important phase

What an SOP really is (and isn't)

The acronym stands for Standard Operating Procedure. The dictionary version is dry and useless. Here's the working definition for our purposes:

★ WORKING DEFINITION

An SOP is a versioned operating guide for a bounded recurring task. It should tell a capable new operator what inputs are allowed, what to do, how to check the result, and when to stop or escalate instead of guessing.

It is not a checklist. It is not "the documentation." It is not a wiki page. Those are subsets of an SOP. An SOP includes context, conditions, variables, examples, and what-to-do-when-things-go-wrong, not just steps.

Here is what an SOP is not:

The independent-review test

Whenever you finish drafting an SOP, run this test in your head:

Imagine a smart human you've never met starts at your company tomorrow morning. They are competent in the general field but have zero context about your brand, your clients, your tools, your standards, your past mistakes. You hand them this single document. You then leave for a 2-week holiday with no Slack.

Could they handle the represented cases, recognize missing or unsafe inputs, and escalate the unfamiliar ones with an audit trail?

If the answer is unclear, improve the guide and test it. No SOP guarantees correct execution in every case.

Early drafts often contain hidden assumptions. Treat the first version as a hypothesis: have another qualified person follow it on representative and edge cases, then revise from observed gaps.

The 7 anatomical parts of an SOP

The following seven sections are a useful starting structure. Adapt them to the workflow, regulatory context, users, and consequence of failure.

①Goal & Definition of Done. What does success look like? How will the assistant (and you) know the task is complete and correct?
②Context & Background. Who is this for? What product, brand, audience? What does the assistant need to know about your world that it can't infer?
③Inputs. What information / files / variables does the task need to start? Where do they come from? What format?
④Procedure. The actual step-by-step. Numbered. Atomic (one action per step). With tools / commands / specifics named.
⑤Conditions & Variables. "If X, do Y. If Z, do W." The branching logic. The if/else. This is the part most SOPs are missing.
⑥Quality bar & Examples. Use synthetic or rights-cleared cases, including boundary, malformed, adversarial, refusal, and failure cases. Authorized real data requires lawful basis, minimization, vendor review, retention, and access controls; anonymization alone is insufficient.
⑦Failure modes & Escalation. What goes wrong commonly? When should the assistant stop and ask you instead of guessing? Who do they ask?

Variables, conditions, and edge cases

Section ⑤ is where amateur SOPs become professional SOPs. A junior writer gives a sequential list. A senior writer gives a decision tree.

Watch the difference:

Junior version:"Pull the daily ad spend report from Meta Ads Manager and send it to the team."
Senior version:
1. Pull yesterday's spend by campaign.
2. Compare to the trailing 7-day average per campaign.
3. If a metric crosses its account-calibrated investigation condition after the required sample and lag, create a review flag.
4. If delivery is unexpectedly zero, verify schedule, status, permissions, tracking and upstream availability; do not guess a cause.
5. If none of the above, format as the "all clear" template.
6. Save a draft to the approved tenant-isolated destination; an authorized owner controls recipients and release time.
7. Edge case: if a campaign has <3 days of data, exclude from the comparison and note it.

The senior version answers questions before they're asked. The assistant doesn't have to guess what to do when the data looks weird. Edge cases are where AI assistants fail loudly — and a good SOP catches them before they happen.

Full example #1 — Operational SOP

Here is a synthetic, non-executable SOP specimen for learning structure. Replace every placeholder through authorized configuration; keep account identifiers, credentials, recipients and private data outside model context. Notice how every one of the 7 sections is present.

SOP — Example Ads Health Review For: [TENANT_FROM_SERVER_CONTEXT] · accounts [SERVER_ENFORCED_ALLOWLIST] Owner: [AUTHORIZED_ROLE] · contact: [APPROVED_ESCALATION_ROUTE] Last reviewed: [YYYY-MM-DD · TIMEZONE] ① GOAL & DEFINITION OF DONE - A schema-validated draft in [TENANT_ISOLATED_DESTINATION] by the configured review window. - Lists every campaign with anomalous spend or performance from yesterday. - "Done" = a qualified owner verifies provenance, completeness and recipients, then explicitly approves or rejects release. ② CONTEXT - Brand X runs always-on Meta campaigns across 3 funnels: prospecting, retargeting, retention. - Sensitive: ad accounts have been restricted twice in the past — never use browser automation. - Audience cares about ROAS and spend pacing more than vanity reach metrics. ③ INPUTS - Only the approved fields from a read-only, least-privilege connector; the model receives no credentials. - The trailing 7-day rolling average per campaign (computed, not pulled). - The tenant/account allowlist is enforced by the service before every retrieval and cannot be changed by prompt text. ④ PROCEDURE 1. Pull yesterday's metrics: spend, ROAS, CTR, impressions, by campaign. 2. Compute trailing 7-day avg for each campaign. 3. Apply documented investigation conditions calibrated from the account baseline, sample, lag, unit economics and response SLA. A flag is not a diagnosis or permission to change delivery. 4. Format the report using the Template (Section ⑥). 5. Save a draft to the approved tenant-isolated review queue. 6. A named owner verifies the complete draft and exact recipients before release; otherwise stop. ⑤ CONDITIONS & VARIABLES - Edge: campaign has <3 days of data → exclude from avg comparison, note it. - Edge: account-level error pulling data → STOP, do not post a partial report. Escalate through [APPROVED_ROUTE]. - Edge: today is a public holiday in target market → mention "Note: holiday in [market] yesterday" in the header. - Variable: currency and timezone come from verified tenant configuration. If conversion is required, record the approved rate source and timestamp; never silently convert. ⑥ QUALITY BAR & EXAMPLES ✅ ACCEPTABLE synthetic example output (cite the source and label assumptions) ❌ BAD example: posting raw numbers without flags / interpretation. The point is the FLAGS, not the data dump. ⑦ FAILURE MODES & ESCALATION - API returns 0 results → the credential broker inspects current metadata and health; escalate through [APPROVED_ROUTE], never expose or repair credentials in model context, and do not silently post empty. - An unusual concentration of flags → escalate for investigation without guessing a market-wide cause. - If anything seems off and you're not sure → STOP and ask. Better to skip a day than post wrong data.

That's it. Roughly 350 words, fully self-contained, no external dependencies in the prose. A new assistant can run it tomorrow morning. So can a human freelancer. That portability is the test of a real SOP.

Full example #2 — Creative SOP

People assume SOPs only work for routine tasks. Wrong. Creative tasks need better SOPs, because the variability is where assistants get lost. Here's one for a creative task — writing Instagram captions on-brand.

SOP — Instagram Caption Writer · Brand X For: every post on @brand_x_official Owner: Mazin ① GOAL & DEFINITION OF DONE - 3 caption variants per post brief. - Each ≤ 220 characters (room for emoji/hashtags). - Sounds like Brand X (see voice rules below). - Includes 1 CTA per caption (link in bio / DM us / shop now). ② CONTEXT — BRAND X VOICE - Tone: warm, expert, slightly playful. Never corporate. Never "marketing-speak." - We talk TO the customer, not AT them. Use "you", not "consumers." - We never use: "elevate", "unleash", "game-changer", emojis > 1 per caption. - We always include: a sensory detail (smell / texture / sight) when the post features a product. - Cultural register: GCC + Egypt audience, English with light Arabic phrases OK ("yalla", "khalas") if it fits. ③ INPUTS - Post brief: image/video description, product featured, campaign theme. - Brand voice file: /assets/brand-x-voice.md (latest version). - Synthetic or specifically authorized, minimized examples with source, context, date, purpose, and retention policy. ④ PROCEDURE 1. Read the brief. 2. Identify the ONE thing the post is about. (If 2+ things, ask for clarification — don't try to combine.) 3. Draft variant A: a "story open" — start with a small concrete observation that leads to the product. 4. Draft variant B: a "direct" — lead with the customer benefit in plain words. 5. Draft variant C: a "playful" — a short hook that uses the visual element of the post. 6. Run each variant through the voice check (Section ⑤). 7. Output as a numbered list, with a 1-sentence explanation under each of why it works. ⑤ CONDITIONS & VARIABLES - IF post is for a launch → all 3 variants must mention "now available" / launch language. - IF post is for an evergreen product → no urgency language, focus on benefit. - IF post is reactive (news / event) → tone shifts to acknowledging-the-moment, never opportunistic. - IF brief mentions price → include in caption B only (the direct one). ⑥ QUALITY BAR & EXAMPLES ✅ GOOD: "Two months of testing this oil and the smell is what got my whole team. Earthy, no perfume, calms a long day. Bottle's small — and that's the point." ❌ BAD: "Discover our revolutionary new oil that will elevate your wellness journey! 🚀✨ Game-changing!" ⑦ FAILURE MODES & ESCALATION - Brief is too vague (e.g., "post about the new product") → ask for product name + 1 sensory detail before drafting. - All 3 variants come out similar → start over with more contrast in approach (story vs direct vs playful). - Anything that sounds AI-generic → rewrite. The check: would I personally send this caption to a friend? If no, scrap it.

Same 7 sections, same shape — but the substance is creative. The "voice rules" subsection is doing a lot of heavy lifting; in real life, that subsection lives in a separate brand-voice.md file referenced from many SOPs.

Common SOP failure patterns

"It's all in my head, I'll just type it up quickly." No. The thing in your head is half-formed. It needs to be externalised before it can be tested. The SOP is the externalisation. Every time you find yourself "knowing" the answer to a question the SOP doesn't address, that's a missing section. Add it.

The "frozen brilliant draft." First-time SOP writers polish v1 to death. Don't. SOPs improve with USE, not with editing. Get to v0.7 and start using it. The errors will tell you what to fix.

"This SOP is too long." Length alone says little. Make the required path scannable, link supporting detail, test findability under realistic conditions, and remove text that does not help a safe decision.

No evaluation examples. Include representative accepted and rejected cases where examples help, plus edge, adversarial, abstention, and escalation cases appropriate to the task. Remove or anonymize sensitive data.

Storing SOPs where you can't find them. A perfect SOP buried in a Notion page no one knows the URL of is functionally a missing SOP. See the "Tools" section below — pick a single home for SOPs and put them all there.

The screen-recording → SOP workflow

The fastest way to draft an SOP is to record yourself doing the task once, then convert the transcript into structure. Workflow:

1Open Loom (or QuickTime). Hit record.
2Do the task once, narrating everything out loud — including what you're checking, why you're skipping that menu, why you're picking that option.
3Transcribe only a synthetic demonstration or a recording with participant consent and approved processing. Evaluate the vendor's access, retention, training, regional, and deletion terms; never upload secrets or unrelated people, screens, or documents.
4Hand the transcript to Claude or ChatGPT with this prompt: "Convert this transcript of me doing a task into an SOP using these 7 sections: Goal, Context, Inputs, Procedure, Conditions, Quality bar with examples, Failure modes. Where you don't have enough info, leave a [TODO] placeholder for me to fill."
5Edit the result. Fill the TODOs. This is the most important step — the AI's draft is your skeleton, not your final.
6Version the specification. Run the independent-review and synthetic eval suite. Iterate.

A recording and transcript may reveal tacit checks that a blank-page draft misses. Compare the resulting SOP's accuracy and total authoring/review time with your normal method before deciding whether the workflow helps.

SOP review checklist

Before you "ship" an SOP into Phase 4 (where you'll match it to an assistant), run it through this checklist:

SOP Readiness Check

Where to write & store SOPs

Pick ONE home for all your SOPs. Stick to it. Don't spread them across Notion, Google Docs, Slack, your Desktop, and "I'll remember." Choose one of these patterns:

★ Markdown files in a folder

Best for technical / dev / Claude Code workflows. Files like sops/daily-ads-report.md. Easy to feed into AI, easy to version-control, easy to back up.

For Claude Code users
Notion pages

Best for teams, non-technical operators, and SOPs that need rich formatting / images. Use a single "SOPs" database with consistent fields. Watch out for fragmentation.

For team-shared SOPs
Google Docs in one folder

Lowest friction, near-universal access, easy to share. Weakness: no structure enforcement, easy to drift. Works well for a solo operator.

Solo & lowest friction

Whichever you pick, a few rules: file names are descriptive (not sop-1.md), every file has a date / version line at the top, and you have a single "index" doc that lists all your SOPs in one place.

From SOP to "skill file" (the bridge to Phase 4)

You'll hear the term "skill file" a lot in the AI world — Claude Skills, OpenAI's GPT instructions, Custom GPT knowledge files, etc. Different vendors, same concept.

★ THE TRANSLATION

A skill file may encode part of an SOP for an AI system, but it also needs platform-specific instructions, scoped tools, input/output contracts, security boundaries, evaluation cases, and operational controls. Treat the platform's current documentation as the source of truth.

What this means: your SOPs are already most of the way to being skill files. You don't have to learn a separate craft. You just need to know how to translate.

A tested SOP gives you useful design material. Phase 4 adds tool selection and boundaries; later phases still require implementation, evaluation, governance, and ongoing operation.

Your task before Phase 4

Pick one of the top-3 tasks you circled in Phase 2. Write an SOP for it.

  1. Use all 7 sections.
  2. Use as much detail as the operator needs, with a scannable main path and linked supporting material.
  3. Use the screen-recording trick if you can.
  4. Run the independent-review test with representative synthetic cases.
  5. Save it in your chosen home (markdown / Notion / Google Doc), in a folder named SOPs/.

Do not move to Phase 4 without a finished first SOP. Reading Phase 4 without a real SOP in hand is like reading a recipe with no ingredients on the counter.

Recap

You now hold the keystone skill of the AI Director. Everything from Phase 4 onward is about putting this skill to work.